GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
iarcuschin
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
iarcuschin
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
The root contains the selected FP32 PEFT adapter and tokenizer. base_model/
contains the complete frozen upstream snapshot at revision b968826d9c46dd6066d109eabc6255188de91218.
Load it in BF16 and keep the adapter in FP32. The adapter is intentionally
unmerged: merging it into BF16 base weights would change the arithmetic that
this experiment measures. The inference helper loads this complete bundle:
bash
python infer.py iarcuschin/gpu-awareness-Qwen3-8BUse the pinned runtime recorded in experiment/source/environment-freeze.txt.
Other hardware or kernels can produce different answers. Only single-token
A/B probabilities and greedy outputs were evaluated; this is not an assessment
of general chat capabilities.
Fresh pretrained initialization; seed 42; rank-16/alpha-32 LoRA on all text attention and MLP projections; dropout disabled; eager attention; one answer per GPU per update; mean of the two FP32 gradients; one AdamW optimizer on RTX PRO 6000. Linear warmup for 100 updates to 5e-5, cosine decay to zero at 1,000 updates; evaluate every 25; stop after eight evaluations without a strictly lower mean loss, with patience counted after warmup. Save each improvement and the periodic 100-update recovery checkpoints.
experiment/ includes the resolved config, full logs, all saved checkpoints,
evaluation logits/hidden states, software provenance and source. manifest.json
records artifact hashes. Model-specific formatting is explicit in the config.
Source and figures: https://github.com/FlyingPumba/gpu-awareness
This derivative uses Qwen/Qwen3-8B.
Its original model card, license and any use-policy files are retained under
base_model/; the upstream terms continue to apply to the model materials.